tranhieutt/llm-app-patterns

LLM Application Patterns

Architecture decision matrix

| Pattern | Use when | Cost | |---|---|---| | Simple RAG | FAQ, docs Q&A | Low | | Hybrid RAG (semantic + BM25) | Mixed query types | Medium | | Function calling | Structured tool use | Low | | ReAct agent | Multi-step reasoning | Medium | | Plan-and-execute | Complex decomposable tasks | High | | Multi-agent | Research, critique-refine | Very High |

RAG: critical config numbers

CHUNK_CONFIG = {
    "chunk_size": 512,       # tokens — sweet spot for most docs
    "chunk_overlap": 50,     # prevents context loss at boundaries
    "separators": ["\n\n", "\n", ". ", " "],
}
# Hybrid search alpha: 1.0=semantic only, 0.0=BM25 only, 0.5=balanced

RAG: retrieval strategies

# Basic: semantic search
results = vector_db.similarity_search(embed(query), top_k=5)

# Better: hybrid (semantic + keyword via RRF)
def hybrid_search(query, alpha=0.5):
    return rrf_merge(vector_db.search(query), bm25_search(query), alpha)

# Best for recall: multi-query (3 variations, deduplicate)
queries = llm.generate_variations(query, n=3)
results = deduplicate([semantic_search(q) for q in queries])

RAG: generation prompt template

RAG_PROMPT = """Answer based ONLY on the context below.
If insufficient, say "I don't have enough information."

Context: {context}
Question: {question}
Answer:"""

Agent: function calling loop

messages = [{"role": "user", "content": question}]
while True:
    response = llm.chat(messages=messages, tools=TOOLS, tool_choice="auto")
    if not response.tool_calls:
        return response.content
    for call in response.tool_calls:
        result = execute_tool(call.name, call.arguments)
        messages.append({"role": "tool", "tool_call_id": call.id, "content": str(result)})

Production: caching (only temperature=0 responses)

def get_or_generate(prompt, model, **kwargs):
    deterministic = kwargs.get("temperature", 1.0) == 0
    if deterministic:
        key = sha256(f"{model}:{prompt}:{json.dumps(kwargs, sort_keys=True)}")
        if cached := redis.get(key): return cached
    response = llm.generate(prompt, model=model, **kwargs)
    if deterministic: redis.setex(key, 3600, response)
    return response

Production: retry + fallback

from tenacity import retry, wait_exponential, stop_after_attempt

@retry(wait=wait_exponential(multiplier=1, min=4, max=60), stop=stop_after_attempt(5))
def call_llm(prompt): return llm.generate(prompt)

# Fallback chain
for model in [primary] + fallbacks:
    try: return llm.generate(prompt, model=model)
    except (RateLimitError, APIError): continue

LLMOps: key metrics

Latency : p50, p99 response time
Quality : satisfaction (thumbs), task completion %, hallucination rate
Cost    : cost_per_request, tokens_per_request, cache_hit_rate
Health  : error_rate, timeout_rate, retry_rate

Embedding model selection

| Model | Dims | Cost | Use | |---|---|---|---| | text-embedding-3-small | 1536 | $0.02/1M | Most cases | | text-embedding-3-large | 3072 | $0.13/1M | High accuracy | | bge-large (local) | 1024 | Free | Self-hosted |